Skip to content

Concepts

Before we work with data, we need to talk about how we think about data. This section introduces nine principles that run through the entire course. They are not statistical methods or bioinformatics tools. They are habits of mind: ways of questioning assumptions, interpreting results, and avoiding the reasoning traps that produce unreliable science.

Stop copying. Start thinking.

The Principles

  • The Inherited Default

    Every threshold and pipeline parameter was a decision made by someone else, in a different context, for different data. Do you know where yours came from?

  • The Correlation Trap

    A significant p-value can appear by chance, be manufactured by selection bias, or reflect a trivially small effect inflated by sample size. It is the beginning of a question, not the answer to one.

  • The Naked Number

    A number without biological context is naked. No statistic has inherent meaning until you know the organism, the environment, and the question.

  • The Fishing Trip

    Finding a pattern is not the same as testing a hypothesis. When you explore data and find something interesting, you have generated a hypothesis, not confirmed one.

  • The Why Layer

    Reporting that two groups differ is an observation. Explaining why they differ is science. Every finding needs a proposed mechanism.

  • The Honest Graph

    A graphic is an argument. Every visual choice determines what the reader sees. An honest graph serves the data. A dishonest graph, even an unintentionally dishonest one, serves the story.

  • The Right Question

    Before you touch your data, write down the question you are actually asking in one sentence. Every analytical decision that follows should be answerable by reference to that question.

  • The Hidden Prior

    A prior is what you believe before you look at the data. A prior that is examined and justified is a hypothesis. A prior that is invisible is a bias.

  • Buttons and Zippers

    A zipper fails catastrophically and opaquely. Buttons are slower but independently understandable and fixable. Your pipeline is a zipper. Your understanding of it should be buttons.

In brief

The Inherited Default — The 97% OTU threshold, the Q=20 quality cutoff, the V3-V4 primer pair: none of these are laws of nature. Before you run your pipeline, you should be able to say where each parameter came from and whether it applies to your data.

The Correlation Trap — In a standard 16S dataset you test hundreds to thousands of taxa simultaneously, with data shaped by primer bias, in cohorts large enough to detect trivial effects, often without a pre-specified hypothesis. All four traps are active at once. Always report effect size alongside p-value.

The Naked Number — FST = 0.68 in bed bugs and FST = 0.099 in cockroaches reflect completely different dispersal biologies. A UniFrac distance, a Shannon index, a PERMANOVA R-squared: none of them mean anything without the biology alongside them.

The Fishing Trip — Running a PCoA, spotting a cluster, and then testing whether that cluster is significant is HARKing. Decide which metadata variable is your primary hypothesis before you look at the plot. If you looked first, say so.

The Why Layer — A significant PERMANOVA tells you that community composition differs between groups. It does not tell you why. The why layer requires you to know your system: the organisms, the environment, and the biological plausibility of the relationship you are describing.

The Honest Graph — The order of taxa in a legend, the colour palette, the grouping of samples, the decision to collapse rare taxa: each is an analytical decision with consequences for what the reader concludes. Always ask whether your visualisation would look different if the result had gone the other way.

The Right Question — Bray-Curtis dissimilarity and UniFrac both measure beta diversity. They measure different things. Using the wrong one for your question is like solving for straight-line distance when the problem required travel distance. The question determines everything.

The Hidden Prior — Higher diversity is always healthier. The microbiome causes the phenotype. Findings generalise across cohorts. These are not wrong beliefs. They are unexamined ones. Before you open your data, write down what you expect to find and where that expectation came from.

Buttons and Zippers — Quality trimming, merging, dereplication, clustering, chimera removal, taxonomic assignment: each step changes your data in a consequential way. Running a pipeline end to end without inspecting each step is a zipper. Understanding each step is buttons.